Oncofusion

start / graph / ana-q146-on-one-universe-with-one-rule-the-lineage-switch-pair-is-indistinguishable-from-an-unrelated-infant

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q146-one-universe-jaccard

On one universe with one rule and a structure-respecting null, the two lineages of one infant overlap no more than two unrelated KMT2A::AFF1 infants do

also called q146 headline · one-universe Jaccard

One baby's leukaemia changed from a B-cell type to a myeloid type, and we asked whether the same genes stayed switched on by the cancer protein. Re-measuring it carefully, the two versions of this child's disease share about as many target genes as two completely different children do.

The same thing, for a scientist

Two earlier beats measured the p279 (infant B-ALL) to p318 (the same infant's lineage-switched AML) overlap of KMT2A::AFF1 oncoprotein-target promoters and disagreed in direction: q142 reported Jaccard 0.117 against 0.149 for two unrelated infants, q144 reported 0.400 against a cohort median of 0.132. Re-running both on a single measurability-defined universe with a single calling rule and an activity-stratified permutation null gives J = 0.328 in the primary cell (universe 10,248 promoters measurable in both samples, 52.8% of the 19,404 refGene promoter set; 4,347 and 3,539 calls after within-pair depth matching). The excess over the structure-respecting null is 1.33x (permutation p = 0.0005), which places the same-patient pair at the 37th percentile of 27 unrelated KMT2A::AFF1 sample pairs and INSIDE the central 80% of that distribution. The pre-registered verdict is NEITHER: the same-patient pair is not distinguishable from an unrelated child, and the direction of the published disagreement flips with the calling rule.

Go deeper
What was actually measured

One universe. U_pair = promoters with nN+nC >= 20 fragments in BOTH members of the pair. This is the only universe that treats an uncovered promoter as unmeasured rather than as not a target. For p279/p318 it is 10,248 of 19,404 promoters (52.8%), and the coverage figure is printed above every rate in the output, never below it.

One rule. The paper's own criterion 1: N/C ratio > 1.75 and binomial upper-tail p < 1e-5, on pooled N1+N2 versus pooled C1+C2 fragment counts, after within-pair depth matching (both libraries binomially thinned to the pair's minimum per-arm total, seed 20260924) so the binomial test has identical power in both members.

One null. An activity-stratified label permutation: promoters binned into deciles of joint coverage sqrt(tot_A * tot_B), sample B's call labels permuted within each decile, 2,000 draws. Both set sizes and the activity distribution of the calls are preserved exactly; only the promoter-identity correspondence is broken.

The numbers, with coverage first
value
universe (measurable in both)10,248 promoters
comparable fraction of 19,40452.8%
p279 calls / p318 calls4,347 / 3,539
intersection / union1,948 / 5,938
Jaccard0.328
independence null (this pair)0.235
activity-stratified null (median)0.247
excess over the structure-respecting null1.33x
permutation p (one-sided)0.0005
unrelated KMT2A::AFF1 pairs (n=27), median excess1.361
same-patient percentile in that distribution37th
inside the central 80% of unrelated pairsyes

There IS excess overlap over chance, and it is highly significant. It is simply the same excess that any two KMT2A::AFF1 children show.

Why this is the pre-registered NEITHER branch and not a q142 win

The pre-registration named three ways to land on NEITHER, and two of them fired.

  1. The direction flips with the rule. Across the 3x2 factorial of {U_pair, U_27, U_full} x {paper absolute rule, q142 quantile rule}, the same-patient pair is BELOW the unrelated median excess in all three absolute-rule cells and ABOVE it in all three quantile-rule cells. Three of six either way, against a pre-registered requirement of four of six.
  2. It sits inside the central 80% of the unrelated distribution in all six cells.

So on the comparison as pre-registered — against the whole KMT2A::AFF1 arm — neither beat is vindicated. A separate, explicitly POST HOC analysis that matches on institution does lean toward q142's direction and is reported in its own entry rather than folded in here.

What this does NOT say

It does not say the two lineages are unrelated. Their promoter-signal rank correlations are high (q142 measured Spearman 0.85 for the KMT2A N-terminal channel and 0.94 for H3K4me3 between these same two samples). It says that the thresholded target set — the currency the paper and this corpus argue in — does not preserve patient identity any better than fusion identity does.

Facts and figures
scriptanalyses/q146-one-universe-jaccard/run146.py
statusrun
result summaryJ(p279,p318)=0.328 in a 10,248-promoter pair universe; excess over an activity-stratified null 1.33x (perm p=0.0005); 37th percentile of 27 unrelated KMT2A::AFF1 pairs; inside the central 80%. Pre-registered verdict NEITHER.
universe promoters10248
comparable fraction0.5281
n calls p2794347
n calls p3183539
jaccard0.3281
null independence0.2351
null stratified0.2468
excess over stratified1.329
perm p one sided0.0005
unrelated aff1 pairs27
unrelated aff1 median excess1.361
headline percentile37
seed20260924
n perm2000
inputs
[
  "C:/Temp/q144mag/counts/*.tsv (133 libraries, GSE252378)",
  "analyses/q146-one-universe-jaccard/PREREGISTRATION.md"
]
How this could bite us back
  • If a patient's target set at diagnosis does not predict the target set after lineage switch any better than another child's does, then a menin-inhibitor target list read off a diagnostic sample is not a personalised list, and treating it as one would misdirect a salvage strategy.
  • The same-epitope antibody negative control calls 2,711 promoters in p279 against 4,347 real N-over-C calls - 62% - so a large minority of any target list built this way is reagent, and a therapeutic decision made on such a list inherits that.
Still open
  1. The comparison is one patient. Two of the three lineage-switching samples the paper profiles have a usable paired partner and only one of those two has a partner that calls any sites at all, so n=1 for this design in the entire public record. What would it take to reach n=5?
  2. Rank correlation between p279 and p318 is high (0.85 on the KMT2A N channel) while thresholded-set Jaccard is at the unrelated base rate. Which of the two is the biologically meaningful quantity for a lineage switch, and is the answer different for a therapeutic target than for a mechanism?
  3. Does the same result hold on the paper's own peak intervals rather than refGene promoters? The interval set is the one divergence this beat did not test, because the paper's peak calls are not in the deposit in a per-sample form.
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The N/C ratio > 1.75 with p < 1e-5 calling rule, and the p279/p318 same-infant lineage-switch pair, are the paper's own; p279 is stated to be a B-ALL that lineage switched in response to ALL-directed chemotherapy.

    MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024
    PMID 39472576DOI
    What this evidence cannot show

    The paper calls sites on its own peak intervals with a CD34+ oncoprotein-negative calibration sample that is NOT deposited in GSE252378; this re-implementation uses TSS+/-1kb refGene promoters and no such calibrator, so absolute counts are not comparable to the paper's 1,692.

    Evidence that argues against this (1)
    • The paper reports 46 oncoprotein-binding sites in 148752 where this promoter-interval re-implementation reports 745 - the absolute count is a property of the interval set, so any claim resting on a count must name its intervals. Janssens DH et al. 2024
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    The fragment-level data underlying every number here are the deposited hg38 fragment BEDs of GSE252378; 27 samples carry all four KMT2A arms.

    What this evidence cannot show

    GEO sample descriptions are case-level and have already been shown by this corpus to mislabel an untreated aliquot as treated. Only the fusion partner, institution and pairing fields were used here, all read directly from !Sample_description.

    Evidence that argues against this (1)
    • Institution was inferred from the free-text description string rather than a structured field; a deposit could in principle describe a sample as being from one hospital while it was sequenced at another. Janssens DH et al. 2024
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    Permutation over a restricted, covariate-matched randomisation space - rather than a uniform independence assumption - is the established way to null genomic-interval overlap, because genome organisation makes intervals non-exchangeable.

    regioneR: an R/Bioconductor package for the association analysis of genomic regions based on permutation tests · Gel B, Diez-Villanueva A, Serra E, Buschbeck M, Peinado MA, Malinverni R · Bioinformatics 32:289-291 2015
    PMID 26424858DOI
    What this evidence cannot show

    regioneR is a general framework; the specific stratification used here (deciles of joint promoter coverage) is this beat's choice and is defended in its own entry rather than imported.

    Evidence that argues against this (1)
    • GAT restricts randomisation on isochore and chromosome rather than on assay coverage; a different restriction gives a different null, and no restriction is uniquely correct. Heger A et al. 2013
Connected to